The Lawyer Who Filed AI-Invented Cases He Never Read

A partner at one of the largest law firms in America signed a court filing that cited nine cases. Eight of them did not exist. They had been invented by the firm's own AI tool, and the partner later admitted he had never read the motion before it went out under his name.

A federal judge sanctioned him and two other lawyers, and imposed a rule with a phrase every professional should sit with: the duty to verify your own work is nondelegable. You cannot hand it to a junior colleague, and you certainly cannot hand it to a machine.

We covered a version of this problem earlier in the series with Deloitte, whose government report contained AI-fabricated sources. This is the same failure, verifying AI output before it ships under your name, but with sharper teeth: here the consequence isn't a refund, it's a court sanctioning a named professional, and a black mark that follows them for years. If your business produces expert work that carries your signature, this is the case to understand.

What happened

The firm was Morgan & Morgan, one of the biggest in the country. Its lawyers were representing a plaintiff in a product-liability case against Walmart in federal court in Wyoming. In drafting a routine motion, a lawyer used the firm's in-house AI research tool to find cases supporting the argument. The tool produced citations. They looked real, proper case names, proper-looking reporter numbers, plausible summaries. They were not real. The AI had hallucinated them.

The filing went out with nine citations, eight of which were fabricated. Walmart's lawyers flagged them. The judge ordered the attorneys to produce copies of the cited cases or explain themselves, and when they couldn't, the sanctions followed: fines totaling several thousand dollars, one lawyer removed from the case, and a written opinion that has since been quoted across the profession.

The judge's reasoning is the part worth keeping. Signing a legal document, he wrote, certifies that the attorney actually read it and made a reasonable inquiry into the law. That responsibility can't be delegated. Blindly relying on someone, or something, else to have done the checking is itself the violation. A fabricated case, he noted, is not law at all, and citing one gives the court nothing.

It wasn't a one-off

The instinct is to write this off as one careless firm. It isn't. Around the same stretch, a prominent regional firm with hundreds of lawyers apologized to a federal court after two of its motions turned out to contain AI-hallucinated citations, a partner had used a general AI tool to find support for a legal point and hadn't verified what it returned. The firm called it an unacceptable lapse in judgment, which it was, and which is exactly the point: careful, credentialed professionals, at serious firms, kept making the same mistake.

This is the tell that the problem is structural, not personal. When highly trained people who face professional sanctions for errors nonetheless repeatedly ship fabricated AI content, the issue isn't that they're careless. It's that AI hallucinations are uniquely designed to slip past a busy expert: they arrive in the exact format the expert expects, wearing all the surface features of legitimate work.

Why professionals fall for it

A hallucinated legal citation is a near-perfect trap. It has the right shape. The case name reads like a real case. The citation format is correct. The one-line summary describes exactly the legal proposition the lawyer was hoping to support. Everything about it confirms what the lawyer already wanted to be true, which is precisely when human scrutiny is weakest.

And the tool never claims to be lying. It presents the fabrication with the same fluent confidence as a real answer, because to the model there's no difference; both are just plausible text. The AI isn't distinguishing between the citations it retrieved and the ones it invented. That distinction only exists to the human, and only if the human actually checks.

This generalizes far beyond law. Any professional using AI to help produce work, an accountant citing a regulation, a consultant quoting a study, a medical expert referencing literature, a financial advisor summarizing a rule, faces the identical trap. The output looks authoritative. The fabrications look identical to the facts. And the professional's name, license, and reputation are what's attached to the result.

The consequence that outlasts the fine

The dollar sanctions in these cases were modest. The lasting damage was reputational, and in the legal example it proved durable in a way worth noting: more than a year after the Wyoming sanction, that same partner was denied permission to appear in a high-profile case in another state, with the court pointing directly to the earlier AI-citation episode as a serious ethical lapse. One unverified filing became a professional liability that followed him into unrelated matters.

That's the real cost structure of this failure. The fine is finite and small. The reputational mark is open-ended. For any business whose product is trust and expertise, that ratio should be terrifying, and clarifying.

How to use AI in expert work without signing your name to fiction

Verify every citation, every source, every specific claim. If the AI gives you a case, a study, a statute, a statistic, open the original and confirm it says what the AI claims. This is non-negotiable and it is the entire safeguard. The lawyers who got sanctioned skipped exactly this step, and it is the only step that would have saved them.

Your signature is a certification, so treat it like one. Signing off on work, formally or informally, means you vouch for it. The judge's word was nondelegable, and it applies to any professional deliverable: if your name is on it, you are certifying you checked it. Never sign, send, or submit AI-assisted work you haven't personally verified.

Never delegate the checking to the thing that produced the work. The AI that wrote the brief cannot be trusted to confirm the brief. Verification has to come from an independent source, the actual case reporter, the real regulation, the primary study, not from asking the same tool whether it was right. It will happily confirm its own fabrication.

Build the check into the workflow, not the good intentions. "Remember to verify" fails under deadline pressure, which is exactly when these errors ship. A real control is a required step: no AI-assisted filing goes out until a named person has confirmed each source against the original. Make it a gate, not a hope.

The real lesson

We build AI into professional workflows, including for firms in law, medicine, and other fields where a signature carries weight and a mistake carries consequences. Used well, AI is a genuine force multiplier for expert work, it drafts, it organizes, it surfaces starting points. What it cannot do is be the final authority on truth. That job stays with the human whose name is on the result.

A lawyer signed a brief he hadn't read, and the AI had filled it with cases that never existed. Strip away the specifics and it's the defining professional risk of this era: the tool is fluent enough to be trusted and wrong often enough to be dangerous, and the only thing standing between those two facts is a human who actually checks. Be that human, or don't put your name on it.